Hi list,

Does anyone have interest in my spending time fully documenting our
vblade-based AoE configuration? Here are the basics:

Drives: 48 Seagate ES.2 7200-rpm disks on two 3ware 9650SE-24M8
controllers, two RAID 6 volumes, four hot spares, and LVM to
partition/stripe it up.

Network: Nortel 10GbE switch for IBM bladecenter.

Server: Four, dual core 2.0ghz AMD opterons, 16gb RAM. NUMA and
realtime kernel options (2.6.24.x). 2x Intel 10Gbase-SR adapters w/
multiqueue kernel support.

Clients: 15 10GbE, mixed Intel (stand alone, mtu of 9000) and Netxen
(bladecenter, max supported mtu of 8000).

Performance (total, on extremely optimized kernel, server): Up to
16Gb/sec (2GB/sec) at about 80% cpu utilization.

Performance (netxen clients, packaged distribution kernel w/ latest
aoe driver): Up to ~3Gb/sec (~380MB/sec) peak average, sporadic bursts
of 3.5Gb/sec (~450MB/sec).

Performance (Intel clients w/o multiqueue, packaged distribution
kernel w/ latest aoe driver): Up to ~5Gb/sec (~630MB/sec) peak.

Performance (Intel client, testing config.. exact mirror of server OS
install w/ latest aoe driver): Hits about 8Gb/sec, no problem, no
effort involved. About ~280,000 packets per second at mtu 9000. Seems
to be the limit of the Intel NIC's and/or switch (per-port) with that
MTU size.

Security layer: VLAN trunking, with the REORDER_HDR flag set on the
server and clients. (Note: Both these network cards offload the
rewriting of vlan tags, so this is a pretty low-cpu consuming task for
us.)

Some random comments:

We had a number of frustrating issues with our configuration
initially, but we've been in production for a couple months now.

Of note, though, is that we did have to bump the AoE buffer count
considerably in vblade to get the netxen clients performing well at
all, so the servers involved would load up the Ethernet port queues on
our bladecenters 10GbE switch (all of these cards support traffic
congestion control features and the switch has huge queues, as well).
They'd also do a lot better, but for some reason, when ext3 updates
filesystem metadata/journal, it refuses to work very well with the odd
block/MTU size.

We also had to tune the heck out of the read ahead and page cache
system on the server, but thats more to do with 3ware controllers and
the small read requests vblade generates than vblade itself.

The largest problem our config has is generating client I/O load. As
it turns out, it's a lot of work to ask for more than a couple hundred
megabytes a second of I/O performance in a way the early Linux 2.6 I/O
schedulers will care about.

Other notes: Latency has been great. CPU usage has been low on
clients. The "sporadic bursts" part above is just that.. we have no
idea why sometimes these cards perform better, the type of work they
are performing does not change when this happens.

The new AIO and socketfilter patch will make our environment a little
more sane even though the vlan isolation thing stops vblade from
seeing other vlan's aoe broadcasts (multiple exports on one vlan
become less painful with the socketfilter patch, AIO let's me relax
the vm.dirty[_background]_ratio tuning a bit), so I'm back into the
mode of thinking about vblade. I'll probably be testing this soon.

Also, is there any interest in people using vblade on 10GbE to add a
command line switch to set the buffer count? I can't imagine we're the
only people who have run into this. We're probably going to write up a
patch to do this, since we're going to export to some clients over
1GbE after we get the N7K up and running.

Offtopic pipe dream/note to the guys at Coraid: If you had one device
like the SR2461 that could do RAID 6+hot spares for the ultra
paranoid, had a web interface (internal or external to the box) for
configuration and disk management somewhere near as feature complete
as 3dm2, did 10GBase-SR (not CX4), etherchannel/802.3ad for
redundancy, LVM (or LVM like) partitioning, VLANs with VLAN trunking,
and could be covered under a support contract, I'd be suggesting we
buy one or two next fiscal year. Far fetched, I know.  But, we're
ramping up to deploy a Cisco Nexus 7000 series switch over the next
few months, in part to deal with 10GbE SAN traffic, and I'm not sold
on FCoE given our awesome AoE setup, and we've got that kind of
solution working locally, and if I can buy it off the shelf it saves
me a lot of time, so... :)

Justin

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft 
Defy all challenges. Microsoft(R) Visual Studio 2008. 
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
_______________________________________________
Aoetools-discuss mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/aoetools-discuss

Reply via email to